Sparse Hyperbolic Convolutional Networks with Enhanced Object Localization via GradCAM Analysis

Vijayavallabh Jayamanikandan, Settur Jithamanyu, Lokesh Kumar Rajulapati, Raghunathan Rengaswamy

Keywords: Hyperbolic geometry, Interpretability, Computer vision

TL;DR: We build Hyperbolic GradCam and analyse activation sparse hyperbolic neural networks.

Abstract: Hyperbolic spaces model hierarchical structures within data. Studies have demonstrated that spatial representations in the hippocampus are structured within hyperbolic spaces to optimize efficiency. We explore the use of hyperbolic convolutional networks with sparsity constraints (L1 and Top-k) and analyze the significance of features in the images for classification tasks using GradCAM. We show that applying sparsity constraints to hyperbolic convolutional networks yields performance comparable to established benchmarks and results in greater interpretability. This work develops sparse hyperbolic representations, enhancing interpretability in AI systems.

Track: Full paper (8 pages excluding references, same as main conference requirements)

Submission Number: 12